Device, method and program for natural language processing
Disclosed is a natural language processing technique according to a neural network of high interpretive ability. One embodiment of the present disclosure relates to an apparatus including a trained neural network into which first natural language text is input and that is trained to output second natural language text and alignment information, the second natural language text being in accordance with a predetermined purpose corresponding to the first natural language text, and the alignment information indicating, for each part of the second natural language text, which part of the first natural language text is a basis of information for generation; and an analyzing unit configured to output, upon input text being input into the trained neural network, a predicted result of output text in accordance with a predetermined purpose, and alignment information indicating, for each part of the predicted result of the output text, which part of the input text is a basis of information for generation.
1. A device comprising:
a storage that stores a trained neural network into which first natural language text is input and that is trained to output second natural language text and alignment information, the second natural language text being in accordance with a predetermined purpose corresponding to the first natural language text, and the alignment information indicating, for each part of the second natural language text, which part of the first natural language text is a basis of information for generation; and
a hardware processor that, when reading and executing the trained neural network:
inputs text into the trained neural network; and
outputs a predicted result of output text in accordance with a predetermined purpose, and alignment information indicating, for each part of the predicted result of the output text, which part of the input text is a basis of information for generation.
2. The device according to claim 1 , wherein the trained neural network, when read and executed by the hardware processor:
converts input text into intermediate states;
generates the predicted result of the output text for each processing unit, sequentially from a first processing unit of the output text, with the intermediate states as input; and
outputs a predicted result of a j-th processing unit and information specifying a processing unit in the input text used in generating the predicted result of the j-th processing unit, using hidden states obtained from converting a predicted result of a j−1-th processing unit of the output text, and the intermediate states output from the encoding unit.
3. A non-transitory computer-readable recording medium having a program embodied therein causing a processor to read and execute the trained neural network of the device according to claim 1 .
4. A device comprising:
a storage that stores a neural network into which first natural language text is input and that outputs second natural language text and alignment information, the second natural language text being in accordance with a predetermined purpose corresponding to the first natural language text, and the alignment information indicating, for each part of the second natural language text, which part of the first natural language text is a basis of information for generation; and
a hardware processor that, when reading and executing the neural network:
inputs text for learning into the neural network;
outputs a predicted result of output text for learning and alignment information, with respect to each of learning data given beforehand consisting of pairs of the input text for learning and correct output text for learning; and
updates each parameter of the neural network, in accordance with a value of a loss function calculated based on the predicted result of the output text for learning and the alignment information.
5. The device according to claim 4 , wherein the loss function is characterized by being calculated such that a value of a loss function becomes smaller in a case of a degree of similarity between a first occurrence frequency and a second occurrence frequency being high, than in a case of the degree of similarity being low, the first occurrence frequency being an occurrence frequency of each vocabulary word of a first natural language used in generating the predicted result of the output text for learning, and the second occurrence frequency being an occurrence frequency of each vocabulary word of the first natural language in the input text for learning.
6. The device according to claim 4 , wherein the neural network, when read and executed by the hardware processor:
converts input text for learning into intermediate states;
generates the predicted result of the output text for learning for each processing unit, sequentially from a first processing unit of the predicted result of the output text for learning, with the intermediate states as input; and
outputs a predicted result of a j-th processing unit and information specifying a processing unit in the input text used in generating the predicted result of the j-th processing unit, using hidden states obtained from converting a predicted result of a j−1-th processing unit of the output text, and the intermediate states.
7. A non-transitory computer-readable recording medium having a program embodied therein causing a processor to read and execute the neural network of claim 4 .
8. A method including:
inputting text into a trained neural network; and
of outputting a predicted result of output text in accordance with a predetermined purpose, and alignment information indicating, for each part of the predicted result of the output text, which part of the input text is a basis of information for generation,
wherein the trained neural network, into which first natural language text is input, is trained to output second natural language text and alignment information, the second natural language text being in accordance with a predetermined purpose corresponding to the first natural language text, and the alignment information indicating, for each part of the second natural language text, which part of the first natural language text is a basis of information for generation.
9. A method including:
inputting into a neural network text for learning, with respect to each of learning data given beforehand consisting of pairs of the input text for learning and correct output text for learning;
outputting a predicted result of output text for learning and alignment information; and
a step of updating each parameter of the neural network, in accordance with a value of a loss function calculated based on the predicted result of the output text for learning and the alignment information,
wherein the neural network, into which first natural language text is input, outputs second natural language text and alignment information, the second natural language text being in accordance with a predetermined purpose corresponding to the first natural language text, and the alignment information indicating, for each part of the second natural language text, which part of the first natural language text is a basis of information for generation.